Influence Functions for CART

This talk deals with measuring the influence of observations on the results obtained with CART classification trees. To define the influence of individuals on the analysis, we use influence functions to propose criterions to measure the sensitivity of the CART analysis and its robustness. The proposals, based on jakknife trees, are organized around two lines: influence on predictions and influence on partitions. In addition, the analysis is extended to the pruned sequences of CART trees to produce a CART specific notion of influence. A numerical example, the well known spam dataset, is presented to illustrate the notions developed throughout the paper. A real dataset relating the administrative classification of cities surrounding Paris, France, to the characteristics of their tax revenues distribution, is finally analyzed using the new influence-based tools.